Halfway to 3D: Ensembling 2.5D and 3D Models for Robust COVID-19 CT Diagnosis
Authors: Tuan-Anh Yang, Bao V. Q. Bui, Chanh-Quang Vo-Van, Truong-Son Hy
Organizations: VNUHCM University of Science, Vietnam National University, Vietnam · Ho Chi Minh University of Technology, Vietnam National University, Vietnam · The University of Alabama at Birmingham, United States
We propose a deep learning framework for COVID-19 detection and disease classification from chest CT scans that integrates both 2.5D and 3D representations to capture complementary slice-level and volumetric information. The 2.5D branch processes multi-view CT slices (axial, coronal, sagittal) using a DINOv3 vision transformer to extract robust visual features, while the 3D branch employs a ResNet-18 architecture to model volumetric context and is pretrained with Variance Risk Extrapolation (VREx) followed by supervised contrastive learning to improve cross-source robustness. Predictions from both branches are combined through logit-level ensemble inference. Experiments on the PHAROS-AIF-MIH benchmark demonstrate the effectiveness of the proposed approach: for binary COVID-19 detection, the ensemble achieves 94.48% accuracy and a 0.9426 Macro F1-score, outperforming both individual models, while for multi-class disease classification the 2.5D DINOv3 model achieves the best performance with 79.35% accuracy and a 0.7497 Macro F1-score. These results highlight the benefit of combining pretrained slice-based representations with volumetric modeling for robust multi-source medical imaging analysis. Code is available at https://github.com/HySonLab/PHAROS-AIF-MIH
Figures & tables
Figure 1: Overview of the proposed method. Axial CT slices are first reconstructed into a normalized 128×128×128 volume through preprocessing. From the reconstructed volume, two complementary representations are learned. The 3D branch processes the full volume using a ResNet-18 architecture trained with Variance Risk Extrapolation (VREx) and supervised contrastive learning to improve cross-domain robustness. In parallel, a 2.5D multi-view branch extracts axial, coronal, and sagittal slices and processes them using a DINOv3 backbone. The predictions from both models are aggregated through an ensemble to obtain the final classification.
Figure 2: From each reconstructed CT volume, we extract slices from three orthogonal anatomical planes: Axial view (original acquisition plane), coronal view, sagittal view.
Set
COVID-19
Non-COVID-19
Total
Train
564
660
1,224
Validation
128
180
308
Total
692
840
1,532
Table 1: Data samples in the Multi-Source COVID-19 Detection Challenge
Set
Healthy
Adeno- carcinoma
Squamous Cell
COVID
Train (F)
100
125
5
100
Train (M)
100
125
79
100
Val (F)
20
25
13
20
Val (M)
20
25
12
20
Total (F)
120
150
18
120
Total (M)
120
150
91
120
Table 2: Data samples in the Fair Disease Diagnosis Challenge. For Task 2, the multi-class classification involves four distinct diagnostic categories: Healthy, Adenocarcinoma, Squamous Cell Carcinoma, and COVID-19. For brevity and visual clarity in our subsequent tables and figures, we map these clinical categories to the following abbreviations: ‘Normal’ (Healthy), ‘A’ (Adenocarcinoma), ‘G’ (Squamous Cell Carcinoma), and ‘COVID’ (COVID-19).
Source
Source 0
Source 1
Source 2
Source 3
F1-score
0.8630
0.8408
0.4828
0.8725
Table 3: Per-source performance for Task 1 using the 3D ResNet-18 model.
Class
Precision
Recall
F1-score
A
0.6901
0.9800
0.8099
G
0.7500
0.1200
0.2069
COVID
0.8462
0.8250
0.8354
Normal
0.8293
0.8500
0.8395
Table 4: Class-wise performance for Task 2 using the 3D ResNet-18 model.
Source
Source 0
Source 1
Source 2
Source3
F1-score
0.9430
0.9431
0.4828
0.9194
Table 5: Per-source performance for Task 1 using the 2.5D DINOv3 model.
Class
Precision
Recall
F1-score
A
0.7241
0.8400
0.7778
G
0.5625
0.3600
0.4390
COVID
0.8333
0.8750
0.8537
Normal
0.8462
0.8250
0.8354
Table 6: Class-wise performance for Task 2 using the 2.5D DINOv3 model.
Model
Accuracy
Macro F1
3D ResNet-18
0.8701
0.7648
2.5D DINOv3
0.9351
0.8221
Ensemble
0.9513
0.8321
Table 7: Performance comparison for Task 1 (COVID-19 detection). Best results are shown in bold.
Model
Source 0
Source 1
Source 2
Source 3
3D ResNet-18
0.8630
0.8408
0.4828
0.8725
2.5D DINOv3
0.9430
0.9431
0.4828
0.9194
Ensemble
0.9659
0.9431
0.5000
0.9194
Table 8: Per-source F1 comparison for Task 1. Best values are shown in bold.
Model
Accuracy
Macro F1
3D ResNet-18
0.7677
0.6677
2.5D DINOv3
0.7677
0.7230
Ensemble
0.7677
0.7229
Table 9: Performance comparison for Task 2 (multi-class classification). Best results are shown in bold.
Class
3D ResNet-18
2.5D DINOv3
Ensemble
A
0.8099
0.7778
0.7778
G
0.2069
0.4390
0.4390
COVID
0.8354
0.8537
0.8537
Normal
0.8395
0.8537
0.8354
Table 10: Class-wise F1 comparison for Task 2. Best values are shown in bold.
Method
Avg
H 1
H 2
H 3
H 4
2.5D
0.741
0.910
0.691
0.493
0.869
3D
0.699
0.825
0.608
0.495
0.868
Ens (0.3)
0.739
0.895
0.673
0.496
0.892
Ens (0.5)
0.751
0.914
0.712
0.495
0.883
Ens (0.7)
0.744
0.912
0.696
0.495
0.873
Table 11: Test results for Task 1 across hospitals.
Method
Avg
Female
Male
2.5D
0.555
0.709
0.402
3D
0.572
0.756
0.388
Ens (0.3)
0.568
0.732
0.403
Ens (0.5)
0.561
0.718
0.403
Ens (0.7)
0.633
0.709
0.557
Table 12: Test results for Task 2 with gender-based evaluation.
College of Computer Science and Artificial Intelligence, Shanghai Key Laboratory of Intelligent Information Processing, Fudan University · University of Oxford · The Hong Kong University of Science and Technology